A Brief Survey on Safety of Large Language Models

Q4 Computer Science

Journal of Computing and Information Technology Pub Date : 2024-07-15 DOI:10.20532/cit.2024.1005778

Zhengjie Gao, Xuanzi Liu, Yuanshuai Lan, Zheng Yang

引用次数: 0

Abstract

Large Language Models (LLMs) have revolutionized Natural Language Processing (NLP) and have been widely adopted in various applications such as machine translation, chatbots, text summarization, and so on. However, the use of LLMs has raised concerns about their potential safety and security risks. In this survey, we explore the safety implications of LLMs, including ethical considerations, hallucination, and prompt injection. We also discuss current research efforts to mitigate these risks and identify areas for future research. Our survey provides a comprehensive overview of the safety concerns related to LLMs, which can help researchers and practitioners in the NLP community develop more safe and ethical applications of LLMs.

查看原文本刊更多论文

大型语言模型安全性简述

大型语言模型（LLM）给自然语言处理（NLP）带来了革命性的变化，并被广泛应用于机器翻译、聊天机器人、文本摘要等各种应用中。然而，LLM 的使用引发了人们对其潜在安全和安保风险的担忧。在本调查中，我们探讨了 LLMs 的安全问题，包括伦理考虑、幻觉和及时注射。我们还讨论了当前为降低这些风险所做的研究工作，并确定了未来的研究领域。我们的调查全面概述了与LLM相关的安全问题，有助于NLP界的研究人员和从业人员开发更安全、更符合道德规范的LLM应用。

本文章由计算机程序翻译，如有差异，请以英文原文为准。

求助全文

约1分钟内获得全文求助全文

来源期刊

Journal of Computing and Information Technology Computer Science-Computer Science (all)

CiteScore

0.60

自引率

0.00%

发文量

审稿时长

26 weeks

期刊介绍： CIT. Journal of Computing and Information Technology is an international peer-reviewed journal covering the area of computing and information technology, i.e. computer science, computer engineering, software engineering, information systems, and information technology. CIT endeavors to publish stimulating accounts of original scientific work, primarily including research papers on both theoretical and practical issues, as well as case studies describing the application and critical evaluation of theory. Surveys and state-of-the-art reports will be considered only exceptionally; proposals for such submissions should be sent to the Editorial Board for scrutiny. Specific areas of interest comprise, but are not restricted to, the following topics: theory of computing, design and analysis of algorithms, numerical and symbolic computing, scientific computing, artificial intelligence, image processing, pattern recognition, computer vision, embedded and real-time systems, operating systems, computer networking, Web technologies, distributed systems, human-computer interaction, technology enhanced learning, multimedia, database systems, data mining, machine learning, knowledge engineering, soft computing systems and network security, computational statistics, computational linguistics, and natural language processing. Special attention is paid to educational, social, legal and managerial aspects of computing and information technology. In this respect CIT fosters the exchange of ideas, experience and knowledge between regions with different technological and cultural background, and in particular developed and developing ones.